RewardAI launches OM-1 robot foundation model trained on human data alone, zero-shot across arms and humanoids

zipengfu · x · 2026-09-21

RewardAI introduced OM-1, its first robot foundation model that zero-shot generalizes to table-top arms, industrial arms, and humanoids. It is learned directly from human manipulation data — no teleop or robot data — with claimed near-human dexterity and multi-robot collaboration.

Zipeng Fu added a control-theory take: attention in robot AI goes to the outer loop (mapping language and images to trajectories) and stops there, but RewardAI argues that smoothly executing a trajectory under changing contact and loads matters just as much. Yunlong likened it to classic cascaded control — a slow outer loop outputs trajectories while a fast inner loop handles low-level dynamics — noting that waypoints alone aren't enough for fast actions, and impedance control is a useful analogy for handling contact.

Original post →

More from Embodied

Embodied channel →